Do LLMs Have a Sense of Time? Zero-Shot Survival Curve Prediction with Frontier Language Models

Abstract

A survival curve S(t) = P(T > t) gives the probability that an event hasn’t happened yet by time t, and it’s the standard framework for modeling time-to-event outcomes in clinical biostatistics. For example, how long until someone is readmitted to the hospital or revisits the ED after discharge. Crucially, survival curves only exist as a population-level construct, never as a per-instance label a model could learn to reproduce, since no individual ever has an observed survival curve of their own. This talk asks whether frontier LLMs can nonetheless generate a valid, personal survival curve zero-shot. We benchmark four frontier models on emergency department revisit and hospital readmission prediction, and find that they not only produce structurally valid curves, but also achieve better discriminative performance and calibration than alternative prediction formulations.

Bio

Emma Chen is a PhD student in Computer Science at Harvard University, co-advised by Professor Vijay Janapa Reddi of the Edge Computing Lab and Professor Pranav Rajpurkar of the Rajpurkar Lab. Her research sits at the intersection of machine learning and clinical medicine, with a focus on building multimodal systems in emergency and acute care, including chest radiographs, physiological waveform data, and electronic health records. Her broader goal is to help translate advances in medical AI into tools that are safe, trustworthy, and genuinely useful at the point of care.